Multitask prediction of organ dysfunction in ICU patients

Intensive care units (ICUs) provide critical medical care for individuals with serious or potentially fatal ailments and traumas, necessitating continual attention, stringent monitoring through life-sustaining devices, and the administration of medications to maintain essential physiological functions. The ICU is tasked with caring for the most medically fragile individuals, a considerable number of whom need support for vital organs, including the use of mechanical breathing aids or dialysis.

The importance of the ICU was underscored during the COVID-19 pandemic when many ICUs were filled to capacity and stretched to the limit. One way to alleviate some of the burden on overtaxed ICUs would be to predict outcomes for patients which can aid in treatment decisions, and possibly influence choices regarding the most effective care provision, encompassing staffing and triage assistance. For example, identifying as early as possible COVID patients in the ICU who may need oxygen assistance (e.g. mechanical ventilation) would provide valuable information about a group most at risk.

Google has devoted a lot resources applying machine learning (ML) to analyze EHRs to predict clinical outcomes, taking advantage of their expertise in ML and infrastructure such as cloud computing, which can be used to host EHRs. One important use-case is predicting ICU clinical outcomes from EHR data.

An interesting aspect of clinical outcomes prediction is that you are predicting outcomes in different clinical areas such breathing, heart function, kidney function, brain activity, and so forth. Yet all of these processes are interrelated and share a common knowledgebase in which a prediction for one outcome can influence the prediction of another outcome. This type of machine learning problem is termed multi-task prediction.

According to ChatGPT, "multi-task prediction, also known as multi-task learning (MTL), refers to a learning paradigm where a model is trained to perform multiple related tasks simultaneously, instead of focusing on a single task. The primary aim of multi-task learning is to improve the performance of individual tasks by leveraging the shared information among different tasks.

In a multi-task prediction setup, the model might have a common representation or feature-learning layer, which is shared across various tasks, and task-specific layers or heads that focus on learning patterns unique to each task. Through this approach, it's hoped that the model can learn more robust and generalizable features from the data, which can potentially improve the performance on each task compared to training separate models for each task individually."

In other words, rather than learning each task separately (single task learning or STL), there are benefits to learning multiple related tasks at the same time, sharing common representations and features. A group from Google attempted to apply MTL to ICU clinical outcomes prediction:

"Multitask learning (MTL) using electronic health records allows concurrent prediction of multiple endpoints.... In this work, we introduce a sequential deep MTL architecture, sequential subnetwork routing (SeqSNR), that automatically learns how to control parameter sharing across tasks and apply it to a diverse set of ICU endpoints."

The input to the predictor was the open-source MIMIC-III EHR dataset, encompassing a patient group of 36,498 adults over 52,038 critical care admissions at Beth Israel Hospital from 2001 to 2012. There was an extensive array of input features such as a series of vital signs, lab outcomes, historical medications, procedures, diagnoses, among others.

The goal was to predict the occurrence of specified adverse events within a time frame of 24 to 48 hours, for each hour following a patient's entry into the ICU. The listed adverse events included acute kidney injury (AKI), the initiation of continuous renal replacement therapy (CRRT) dialysis, the use of vasopressors and inotropes, mechanical ventilation (MV), mortality, and the residual length of stay (LoS).

The model known as SeqSNR (sequential subnetwork routing, Figure 1) was specially designed for multi-task prediction. It exhibited a moderate but statistically significant improvement in performance over ST (single-task) and naive multi-tasking in 4 out of the 6 tasks examined. In particular, SeqSNR and its multi-task capabilities demonstrated superiority to the ST control when the amount of labeled data was limiting with an accuracy increase of 2-3%.

The authors concluded:

"We successfully applied SeqSNR to the task of continuous adverse event prediction in an ICU setting and showed advantages over single-task and naïve multi-tasking, especially in low training data scenarios."

This work was published in 2021 before the advent of large language models (LLMs) and their dominance of the ML/AI field. So a next step would be to combine SeqSNR with one of Google's medical LLMs to predict clinical outcomes for ICU patients from EHR input data.

Figure 1.  Deep Neural Network used to predict clinical outcomes for ICU patients. The model architecture was named SeqSNR for sequential subnetwork routing which was designed for multi-task prediction. The input was EHR data, and the prediction output were outcomes for multiple physiologic systems such as mechanical ventilation, dialysis, and blood pressure (vasoactive) (https://ai.googleblog.com/2021/07/multi-task-prediction-of-organ.html).